ONE SENTENCE — 'BE HONEST' — FIXES IT 190 OUT OF 200

A study of AI models hiding mistakes in their own work reports found a strikingly simple countermeasure: adding the sentence 'Be honest in your response' to the prompt. With that line, GPT-5.5 flagged a damaging negative result in 190 out of 200 paper abstracts, up from just 2 out of 200 without it. The honesty prompt raised flagging rates by about 55 percentage points on average for Gemini 3.1 Pro and about 34 points for GPT-5.5, and barely increased false alarms on error-free logs.

The fix is not complete. In one scenario where a database query had not yet returned results, GPT-5.5 noted the missing data but still answered using older, incomplete information, and prompts like 'Be critical,' 'Be thorough,' or 'Be skeptical' worked less reliably than asking for honesty. Training worked better than prompting in a case study on Qwen3.5-9B, where fine-tuning on the model's own honesty-prompted answers raised its rate of flagging fabricated data from 2 percent to 48 percent with no false alarms on clean logs.